A Static-to-Temporal Framework for Interpretable Camera Lens Soiling Severity Estimation in Autonomous Driving.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Sensor Technology

Background:

• Camera lens soiling significantly impairs visual perception in autonomous vehicles, necessitating accurate severity estimation for safety.
• Existing methods lack robustness due to insufficient consideration of opacity, spatial importance, and temporal dynamics in video streams.

Purpose of the Study:

• To develop a robust static-to-temporal framework for camera-soiling severity estimation.
• To address limitations in current methods regarding opacity, spatial importance, and temporal stability.
• To enhance camera-health monitoring and downstream perception safety in autonomous driving.

Main Methods:

• Proposed a Structured Dual-Head Static Model (SDSM) for joint tile-level and image-level soiling prediction with an explicit Structured Severity Score.
• Developed a Two-Stage Stable Diffusion (TS-SD) pipeline to generate SD-Seq data for mechanism supervision, overcoming real temporal data scarcity.
• Introduced a structure-constrained adaptive EMA Module to enhance temporal stability while maintaining the single-frame severity scale.

Main Results:

• Achieved strong cross-domain severity estimation performance on WoodScape, External Test, and OccNuScenes-Dirt datasets.
• Attained a cluster-level Spearman correlation of 0.7876 on the External Test dataset.
• The temporal module reduced Jitter (MAD) by 51.5%, significantly improving temporal stability.

Conclusions:

• The proposed framework offers an interpretable, cross-domain, and deployment-friendly solution for camera-soiling assessment.
• This approach enhances the reliability of visual perception for autonomous driving systems.
• The method effectively addresses the challenges of opacity, spatial importance, and temporal dynamics in soiling estimation.